File size: 2,244 Bytes
1d300e4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ab6ef2e
1d300e4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
---
tags:
  - mic
  - test-artifact
base_model: facebook/bart-base
---

# bart-base-mic

Slim model weights, executable usage and framework expectations in one repository.

- [Contract manifest](integration.json)
- [Transformers usage](integrations/transformers/usage.md)
- [Transformers expectations](integrations/transformers/expectations)
- [Fixture provenance](build.json)
- [How the fixture was shrunk](https://huggingface-mic-mic-coverage.static.hf.space/shrink/bart-base-mic.html)

## Contract

### BART encoder-decoder

Runnable usage and the framework expectations directory are linked above.

## Slim model

### Slim fixture: bart-base-mic

A 3.5M-parameter fixture for the [BART MIC](https://huggingface.co/huggingface-mic/bart-base-mic), covering [`facebook/bart-base`](https://huggingface.co/facebook/bart-base). It is a test artifact, not a usable model, tagged `test-artifact` to keep it out of candidate-model search.

- **Fixture key `c89a9e951b4f`.** It covers every checkpoint whose processor files are byte-identical and whose config matches after shrinking.
- **Same architecture, smaller config:** `BartForConditionalGeneration`, shrunk by the generic rule in [`sweep.py`](https://github.com/huggingface/transformers-mic/blob/main/experiments/sweep.py) (encoder_layers 2, decoder_layers 2).
- **Processor files copied byte for byte** from `facebook/bart-base`: `merges.txt`, `tokenizer.json`, `vocab.json`.
- **Overfit to the contract's training step** (60 steps). Random weights failed the input-sensitivity gate: `input_ids:reverse` moved the logits by only 0.00087 (threshold 0.001).
- **Same tensor names as the real checkpoint.** The real checkpoint has every parameter, and so does the fixture.
- **Saved in float32**, the dtype `facebook/bart-base` loads in by default.
- **Finite in float32, bfloat16, and float16** (forward, loss, gradients).

Built by [`build_fixtures.py`](https://github.com/huggingface/transformers-mic/blob/main/example/build_fixtures.py); [`build.json`](build.json) records the pinned sources, gate results, and file hashes. This repository includes the slim weights, usage and framework expectations; rebuild the other files with `python example/build_fixtures.py --fixture bart-base-mic`.